Robust model-based steady-state feedback optimization for chemical plants

Model-based real-time optimization (RTO) of steady-state operations of chemical plants is of industrial interest due to the fact that the model-based approach can deal with large-scale optimization systems effectively and the inevitable model-plant mismatch can be handled by adapting the model parameters repeatedly. The process operation is continuously improved by on-line computation of the optimal setpoints to be tracked by the lower level control system. It is noted that model updating usually involves a cumbersome re-optimization procedure to obtain the model parameters and may not necessarily improve the model. Thus, in this dissertation, a feedback optimization methodology which does not require model updating is proposed as a complementary approach to RTO. The proposed methodology is based on an extension of a method introduced in optimization-based run-to-run control for batch processes, in which any gradient-based optimization algorithm can be utilized. The plant operation is an integral part of the optimization system to provide measurement data as feedback used in the gradient computation. Thus the proposed methodology is inherently robust to model error and it improves plant operations gradually (iteratively) with successive search directions. Both implicit algebraic models and differential models (describing steady-state spatial distribution for distributed parameter systems) can be utilized directly for gradient formulation through the reduced gradient method and the adjoint method, respectively. The measurement noise effect on the performance of the methodology is quantitatively analyzed. According to the anticipated noise size for the sensors, a performance uncertainty region in the input space is determined to predict the area in which improvement directions are not possible to compute. For computational efficiency and convergence to global optimum, only multi-input single output (MISO) implicit algebraic models are considered in this off-line analysis. Finally, a statistical method for optimization results analysis is presented to on-line reduce the noise effect by evaluating the variability in the optimization variables. Only the optimization results that represent meaningful changes are transmitted as setpoints to the lower process control level. Thus unnecessary and profitless corrective actions caused by stationary measurement noise can be avoided.